The Workload Process: Modeling, Inference and Applications: Michael Reich

Introduction Relationship Between Service Time and Patience Workload and Oered-Load Empirical Results Future Research
The Workload Process:

Modeling, Inference and Applications
Michael Reich
Technion - Israel Institute of Technology
January 2, 2010
Advisors: Prof. Avishai Mandelbaum, Prof. Yaacov Ritov
Outline
1
Introduction
Motivation
Research Outline
Related Work
2
Relationship Between Service Time and Patience
Denitions
The Model
Testing the Relationship
3
Workload and Oered-Load
Denitions
The Workload and Oered-Load of M
t
/GI /N
t
+ GI
Estimation of the Workload and Oered-Load
4
Empirical Results
Stang and Performance Measures
Analysis of the Relationship
U.S. Bank Case Study
5
Future Research
Motivation
Standard assumption in modeling service systems: The
service-time and the patience of customers are independent
Is this assumption valid ?
I can pay my bills on another occasion
I must immediately consult my broker in order to protect my
equities prole
Motivation
Standard assumption in modeling service systems: The
service-time and the patience of customers are independent
Is this assumption valid ?
I can pay my bills on another occasion
I must immediately consult my broker in order to protect my
equities prole
Motivation
Various performance measures are inuenced by the presence of
dependency between patience and service-time
Staffing
Level
Required
Waiting
Times
Quality of Service
Patience Staffing Rule
Abandons
Service
Times of
Served
Customers
Relationship between
Patience-Service Time
Motivation
Mean Service-Time as a Function of Waiting-Time
U.S. Bank - Retail Banking Service - Weekdays - January-June, 2006
190
210
230
250
270
290
150
170
190
210
230
250
270
290
0 50 100 150 200 250 300
Waiting Time
Fitted Spline Curve E(S|>W=w)
Research Outline
A model for the relationship between patience and service-time
Denitions of the workload process and the oered-load
functions
Estimation and prediction of the workload/oered-load
Performance analysis of both real and simulated data
Related Work
Eick S.G., Massey W.A., and Whitt W.,
The physics of the Mt/G/ queue. 1993.
Zeltyn S. and Mandelbaum A.
Call centers with impatient customers: Exact analysis and many-server
asymptotics of the M/M/n + G queue. 2005.
Feldman Z., Mandelbaum A., Massey W. A. and Whitt W.
Stang of time-varying queues to achieve time-stable performance. 2007.
Weinberg J., Brown L.D. and Stroud J.R.
Bayesian forecasting of an inhomogeneous Poisson process with
applications to call center data. 2007.
Shen H.
On modeling and forecasting time series of smooth curves. 2008.
Goldberg Y., Ritov Y. and Mandelbaum A.
The best linear unbiased estimator for continuation of a function. 2010.
Denitions
The M
t
/GI
/n
t
+ GI Queue:
M
t
- Nonhomogeneous (time-dependent) Poisson arrivals
GI
- General service-time distribution; May depend on

patience
n
t
- The number of telephone agents over time
GI - General patience distribution
The parameters are assumed to be uninuenced by systems state
Denitions
Notations:
S - Service-time
- (Im)patience
V - Virtual waiting-time, namely the time a customer is
required to wait before entering service
W - Waiting-time of a customer, calculated over both served
and abandoning customers
Note: W = min{V, }
Denitions
Assumptions:
The pair (, S) is independent of V
For those who abandon, we observe V censored by their
(im)patience
For those who are served, we observe V
W is observable for all customers
S is observable only for customers who have > V
Denitions
Biased Sampling:
In common applications, service-time is measured over all
served customers
The prevalent estimator for mean service-time is actually
E(S| > W)
One tends to observes more customers with longer
patience
What if the patience and service-times (associated with the same
customer) are positively correlated ?
Denitions
Biased Sampling:
In common applications, service-time is measured over all
served customers
The prevalent estimator for mean service-time is actually
E(S| > W)
One tends to observes more customers with longer
patience
What if the patience and service-times (associated with the same
customer) are positively correlated ?
The Model
Consider the mean service-time of customers who were served after
waiting exactly w time units:
g(w) = E(S| > W, W = w) = E(S| > w, V = w) = E(S| > w)
Calculations show that
g(w) =
_
u=w
f
(u) E(S| = u) du
_
u=w
f
(u) du
or alternatively,
E(S| = w) = g(w)
g
(w)
h
(w)
,
where f
(w) and h
(w) are the cdf and the hazard-rate function of

the patience,
Examples: E(S) , E(S| = w) , g(w)
Let exp(
8
3600
)
(a) E(S| = w) = 216 (1.2 e
w
21
3600
)
0
50
100
150
200
250
300
350
400
450
500
0 200 400 600 800 1,000
Waiting Time
E(S) E(S|T=w) g(w)
(b) E(S| = w) = 216 (1.2 e
w
21
3600
)
0
50
100
150
200
250
300
350
0 200 400 600 800 1,000
Waiting Time
E(S) E(S|T=w) g(w)
(c) E(S| = w) = 216 + 108 sin(
36
3600
w)
0
50
100
150
200
250
300
350
0 200 400 600 800 1,000
Waiting Time
E(S) E(S|T=w) g(w)
(d) E(S| = w) = 216 + 108 sin(
36
3600
w + 2)
0
50
100
150
200
250
300
350
0 200 400 600 800 1,000
Waiting Time
E(S) E(S|T=w) g(w)
Propositions
The expectation of service-time equals the expectation of the
service-time given no-waiting:
E(S) = E(S|W = 0)
For a xed w, w 0
1
g(w) > E(S| = w) g
(w) > 0
2
g(w) < E(S| = w) g
(w) < 0
3
g(w) = E(S| = w) g
(w) = 0
Propositions
Assume that E(S| = w) is continuous:
1
If E(S| = w) is an increasing (decreasing) monotone
function, then g(w) is also an increasing (decreasing)
monotone function; The opposite is not necessarily true
2
g(w) is an increasing (decreasing) monotone function if and
only if g(w) dominates (is dominated by) E(S| = w)
3
g(w) converges to a limit L if and only if E(S| = w)
converges to L,
L = lim
w
g(w) = lim
w
E(S| = w)
Propositions
Our model can be applied to any transformation T on S which
satises:
1
E(|T(S)|) <
2
E(T(S)| = w) is continuous with respect to w
E(T(S)| = w) = E(T(S)| > w)
w
E(T(S)| > w)
h
(w)
, w > 0
In particular:
1
E(S
k
| = w) = E(S
k
| > w)
w
E(S
k
| > w)
h
(w)
, w > 0
2
P(S s| = w) = P(S s| > w)
w
P(S s| > w)
h
(w)
, w > 0
We present a statistical test for the relationship between patience
and service-time:
H
0
: E(S| = w) = E(S), w 0
H
1
: Otherwise
Since E(S| = w) is not observable, the test is based on
g(w) = E(S| > W, W = w)
Under the null hypothesis:
g(w) = E(S| = w)
E(S| = w) is constant if and only if g(w) is constant
Let g(W) be a random variable which takes the value g(w)
according to the density function f
W|>W
(w)
We test if the variance of g(W) can be assumed to be zero:
Var (g(W)) = E(g
2
(W)) E
2
(g(W))
Consequently, we choose the statistic for our test to be
T =
_

u=0
g
2
(u) f
W|>W
(u) du
_ _

u=0
g(u) f
W|>W
(u) du
_
2
Construct a permutation distribution for the test statistic:
1
Data
Take all the observations of served customers with strictly
positive waiting-time
2
Discretization
Divide the observations into several groups of a similar size,
according to the ranking of their waiting-times
Calculate the probability of an observation to belong to each
group
3
Permutations
Generate a large number (say 4,000) of permutations by
randomly pairing between the waiting-times and the
service-times
For each permutation, calculate the statistics value
Denote:
t
1
, t
2
, . . . , t
K
- The values of the test statistic in any of the
random pairing permutations
t
- The original samples statistic

The p-value is then approximated by the proportion of samples
with the value of this statistic larger than the original samples
statistic. Explicitly:
p value
K
i =1
1I
{t
i
>t
}
K
Remark:
_
u=0
g(u) f
W|>W
(u) du is equal in all the
permutations, since it is simply the average over all the
observations Thus, a simplied version of the test statistic is the
empirical second moment:
T =
_

u=0
g
2
(u) f
W|>W
(u) du
Denitions
Introduce the following denitions:
1
Resource-k Workload - A stochastic process, representing
the amount of work being processed by resource k at time t,
under the assumptions of innitely many resources of type k,
and that a task that reaches resource k enters service
immediately upon arrival
2
Resource-k Oered-Load - A function of time t 0,
representing the average of Resource-k Workload at time t
t
/GI /N
t
+GI
In order to describe the workload process formally, we introduce
the following notations:
0 = t
0
< t
1
< t
2
< . . . - Increasing sequence of times, where
t
i
represents the arrival time of the i
th
customer
S
i
- Service-time of the i
th
customer, i = 1, 2, . . .
A = {A(t), t 0} - Arrival process; We dene A(0) = 0;
Then A(t) = max{n; t
n
t}
The Workload Process, L = {L(t), t 0}, is then dened by
L(t) =
_
t
0
1I
{S
u
>tt
u
}
dA(u) =
A(t)
i =1
1I
{S
i
>tt
i
}
t
/GI /N
t
+GI
Theorem:
For any time t > 0, L(t) has a Poisson distribution with mean
R(t) = E[L(t)] = E[(t S
e
)] E[S] = E
__
t
tS
(u)du
_
=
=
_
t
(u) [1 G(t u)]du,

where
S is a generic service-time
S
e
is a generic excess service-time
R(t) is by denition the Oered-Load function
t
/GI /N
t
+GI
Insights:
The oered-load lags after the arrival rate -
R(t) = E[L(t)] = E[(t S
e
)] E[S]
Arrival-Rate and Oered-Load
Emergency Depatment of an Israeli Hospital - Sundays
10
15
20
25
30
35
40
45
10
15
20
25
0
5
10
15
20
25
30
35
40
45
0
5
10
15
20
25
00:00 02:00 04:00 06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00
Arrival Rate Offered-Load (Right)
t
/GI /N
t
+GI
Insights:
Alternative Poisson process view:
L(t) =
A(t)
i =1
1I
{S
i
>tt
i
}
Dene a two parameters stochastic process
{A(s, t) =
A(s)
i =1
1I
{S
i
>tt
i
}
, 0 s t}
A(s, t) represents the number of arrivals until time s, which
are still in service at time t
For any pre-specied time t, A(s, t) is a Poisson process in s,
with rate (u) [1 G(t u)]
The expected number of customers at time t, who arrived
during the time interval [s
1
, s
2
], is
_
s
2
s
1
(u) [1 G(t u)]du
t
/GI /N
t
+GI
Insights:
A generalization to the M
t
/GI
t
/N
t
+ GI queue:
Let G
u
be the cdf of the service time of a customer who arrived at
time u
Then,
R(t) =
_
t
(u) [1 G
u
(t u)]du, t 0
M
t
/GI/ queue:
1
The workload process, L = {L(t), t 0}, equals the number
of customers in service (L(t) at time t)
2
The oered-load, R, is estimated as the average number of
customers in service (over all available periods)
M
t
/GI/n
t
queue:
1
Eliminate the customers waiting-times and shift their service
period to start right upon arrival
2
Then, follow the procedure of the M
t
/GI / queue
M
t
/GI/n
t
+GI queue:
1
For the workload process, impute the service-times of
abandoning customers and follow the M
t
/GI /n
t
queue
2
We propose a method to estimate the oered-load, based on
the expression R(t) =
_
t
(u) [1 G(t u)]du

Denitions:
0 = t
0
< t
1
< t
2
< . . . - Time sequence that determines our
estimate for the arrival rate in any time interval
[t
i 1
, t
i
), i = 1, 2, . . .
Assume:
For any i 1, t
i
t
i 1
= d.
The arrival rate is nearly constant over any time interval, but
may dier accross time intervals
Denote
i
as the arrival rate (per d time units) in interval
[t
i 1
, t
i
)
Denitions:
0 = s
0
< s
1
< s
2
< . . . - Service-time sequence, which we use
to estimate G
c
(s
k
) = 1 G(s
k
), k = 1, 2, . . .
Assume:
for any k 1, s
k
s
k1
= r
d = r n, n N
G
c
is approximately linear in any of the time intervals
The proportion of arrivals from time t s
k
, that are still in service
at time t, is G
c
(t s
k
)
The arrival rate in [t s
k
, t s
k1
) is:
k
(t) =
i
k
n
per d time units

or
k
(t)
r
d
per r time units
The rate of arrivals from the time interval [t s
k
, t s
k1
), that
remain in service after time t, is approximately
k
(t) =
k
(t)
r
d

G
c
(t s
k
) + G
c
(t s
k1
)
2
This yields that R(t)
j :s
j
=t
k=1
k
(t)
Assume that:
The arrival rate is bounded by the value M
There exists a value K such that M
_

K
[1 G(u)]du < ,
for any > 0
The empty-system assumption at time 0 can be relaxed for t > K.
Then R(t)
j :s
j
=K
k=1
k
(t) , for any t > K
Remark:
In call centers or service over the counter, K can be taken to
be a few hours
In other type of service systems, such as hospitals, it might be
few days or even weeks
Emergency Department of an Israeli Hospital
aa
Service-time is assumed
homogenous for all time intervals
Service-time is assumed
heterogenous accross time
intervals (but homogenous over
weekdays)
15
20
25
30
35
40
45
0
5
10
15
20
25
30
35
40
45
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
Sunday Monday Tuesday Wednesday Thursday Friday Saturday
Average Load Calculated average Load
15
20
25
30
35
40
45
0
5
10
15
20
25
30
35
40
45
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
0
:
0
0
4
:
0
0
8
:
0
0
1
2
:
0
0
1
6
:
0
0
2
0
:
0
0
Sunday Monday Tuesday Wednesday Thursday Friday Saturday
Average Load Calculated average Load
M
t
/GI
/n
t
+GI queue:
1
For the workload process
Estimate the distribution of the conditional service-time,
S| = t
Impute the service-times of abandoning customers
Follow the M
t
/GI /n
t
procedure
2
For the oered-load
Estimate the marginal service-time distribution from
non-waiting customers
Apply the M
t
/GI
/n
t
+ GI estimation procedure with this
service-time distribution
Iterative Stang Algorithm (ISA), a simulation code developed
by Feldman et al. [07]
Determines time-dependence stang levels aiming to achieve
time-stable delay probability (hence time-stable performance)
In our implementation, we added the feature of dening the
relationship between patience and service-time in the
time-varying M
t
/GI
/n
t
+ GI queue
Remark: In this part of the thesis, we analyze only queues with
homogenous Poisson arrival rate
ISA was applied to three types of M/M
/n
t
+M queueing systems:
Arrivals are according to a homogeneous Poisson process with
arrival rate of 100 customers per time unit
Patience is exponentially distributed with mean 1 time unit
Mean service-time (unconditional) is equal to 1 time unit
Service time, conditional on the patience of a customer, is
exponentially distributed
Relationship between patience and service-time diers accross
models
All performance measures are calculated as an average of 5000
replications
Description of the Relationship:
Increasing Monotone Function E(S| = w) g(w)
No Relation 1 1
Increasing Monotone Function 1.2 e
4w
1.2 0.2 e
4w
Decreasing Monotone Function 0.8 + e
4w
0.8 + 0.2 e
4w
(a) Increasing Monotone Function
0.8
1
1.2
1.4
0
0.2
0.4
0.6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
E(S) E(S|=w) g(w)
(b) Decreasing Monotone Function
1
1.2
1.4
1.6
1.8
2
0
0.2
0.4
0.6
0.8
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
E(S) E(S|=w) g(w)
aa
Comparison between the mean
service-times of served customers
as a function of the stang level
Comparison between the mean
service-times of served customers
as a function of the probability of
waiting
0.98
1.00
1.02
1.04
1.06
1.08
1.10
1.12
E
(
S
|
>
W
)
0.90
0.92
0.94
0.96
0.98
70 80 90 100 110 120
E
(
S
|
Staffing
Increasing Monotone Decreasing Monotone Constant
0.98
1.00
1.02
1.04
1.06
1.08
1.10
1.12
E
(
S
|
>
W
)
0.90
0.92
0.94
0.96
0.98
0 0.2 0.4 0.6 0.8 1
E
(
S
|
P(Wait>0)
Analysis of the Square-Root Stang Rule:
The rule for M/M/n+M queue is given by
n = R +
R ,
where,
R is the oered load
is the Quality of Service (QoS) parameter, determined by
the Garnett function:
=
_
1 +

h(
_
/)
h()
_
1
, < < ,
where is the required probability of delay
For any target , we ran an ISA simulation for each model
Dene the implied quality of service grade
ISA
n
ISA
R
R
,
where n
ISA
denotes the result stang level of the simulation
0.4
0.5
0.6
0.7
0.8
0.9
1.0
D
e
l
a
y

P
r
o
b
a
b
i
l
i
t
y
0.0
0.1
0.2
0.3
0.4
-3 -2 -1 0 1 2 3
D
e
l
a
y

P
r
o
b
a
b
i
l
i
t
y
Beta
Garnett Function Increasing Monotone Constant Decreasing Monotone
In fact, the mean service-time that the system faces, S
, (due to
served customers) is dierent from the unconditional mean
service-time and is aected by the quality of service (determined
by stang level, N)
We dene a modied oered-load expression, given by:
R
= E(S
(N))
(a) Increasing Monotone Function
0.5
0.6
0.7
0.8
0.9
1.0
D
e
l
a
y

P
r
o
b
a
b
i
l
i
t
y
0.0
0.1
0.2
0.3
0.4
-3 -2 -1 0 1 2 3
D
e
l
a
y

P
r
o
b
a
b
i
l
i
t
y
Beta
Garnett Function R=E(S) R=E(S|>W)
(b) Decreasing Monotone Function
0.5
0.6
0.7
0.8
0.9
1.0
D
e
l
a
y

P
r
o
b
a
b
i
l
i
t
y
0.0
0.1
0.2
0.3
0.4
-3 -2 -1 0 1 2 3
D
e
l
a
y

P
r
o
b
a
b
i
l
i
t
y
Beta
Garnett Function R=E(S) R=E(S|>W)
Notice that the oered load is not inuenced by the relationship
Many other performance measures are inuenced
(a) P(Ab) vs P(Wait)
0.08
0.10
0.12
0.14
0.16
0.18
0.20
P
(
A
b
a
n
d
o
n
)
0.00
0.02
0.04
0.06
0.08
0 0.2 0.4 0.6 0.8 1
P
(
A
b
a
n
d
o
n
)
P(Wait>0)
(b) E(Wait) vs P(Wait)
0.08
0.10
0.12
0.14
0.16
0.18
0.20
E
(
W
a
i
t
)
0.00
0.02
0.04
0.06
0.08
0 0.2 0.4 0.6 0.8 1
E
(
W
a
i
t
)
P(Wait)
(c) P(Ab) vs E(Wait)
0.08
0.10
0.12
0.14
0.16
0.18
0.20
P
(
A
b
a
n
d
o
n
)
0.00
0.02
0.04
0.06
0.08
0 0.05 0.1 0.15 0.2
P
(
A
b
a
n
d
o
n
)
E(Wait)
(d) P(Ab) vs E(Wait|Wait > 0)
0.1
0.1
0.1
0.1
0.2
0.2
0.2
P
(
A
b
a
n
d
o
n
)
0.0
0.0
0.0
0.1
0.1
0 0.05 0.1 0.15 0.2
P
(
A
b
a
n
d
o
n
)
E(Wait|Wait>0)
Analysis of the Service-Time vs Patience Relationship
Consider a simulation model:
The service-time of a customer with patience = t is a
Log-Normal random variable, S| = t LogNorm((t),
2
),
with pdf
f
S|
(s|t) =
1
s
2
2
e
(ln s(t))
2
2
2
, s, t > 0
Then,
E(S| = t) = e
(t)+
2
2
E(S
2
| = t) = e
2
e
(t)+
2
2
= e
2
E
2
(S| = t)
Var (S| = t) = (e
2
1)e
(t)+
2
2
= (e
2
1)E
2
(S| = t)
Assume that:
is exponentially distributed, with mean
1
E(S| = t) be of the form:

E(S| = t) = a (b e
t()
), > , a > 0, b > 1
The analysis covers the following:
E(S| = w) = 230 (1.2 e
w(
29
3600
8
3600
)
)
=
8
3600
Four values for are considered
A comparison between Var (S| = w) and Var (S| > W = w)
with dierent values of
(a) = 0.01
3000
4000
5000
0
1000
2000
0 200 400 600 800
Var(S|T=t) Var(S|T>W=t)
(b) = 0.1
3000
4000
5000
6000
0
1000
2000
3000
0 200 400 600 800
(c) = 0.5
15000
20000
25000
0
5000
10000
0 200 400 600 800
(d) = 0.8
40000
50000
60000
70000
80000
0
10000
20000
30000
40000
0 200 400 600 800
95% percent condence intervals of the spline estimator for
g(w) = E(S| > W = w)
(a) = 0.01
230
235
240
245
250
210
215
220
225
230
0 20 40 60 80 100 120 140
Mean 95% Confidence Interval Theoretical
(b) = 0.1
230
235
240
245
250
210
215
220
225
230
0 20 40 60 80 100 120 140
(c) = 0.5
230
235
240
245
250
210
215
220
225
230
0 20 40 60 80 100 120 140
(d) = 0.8
230
235
240
245
250
210
215
220
225
230
0 20 40 60 80 100 120 140
95% percent condence intervals of the derivative of the spline
estimator for g(w)
(a) = 0.01
0.2
0.4
0.6
-0.4
-0.2
0
0 20 40 60 80 100 120 140
(b) = 0.1
0.2
0.4
0.6
-0.4
-0.2
0
0 20 40 60 80 100 120 140
(c) = 0.5
0.2
0.4
0.6
-0.4
-0.2
0
0 20 40 60 80 100 120 140
(d) = 0.8
0.2
0.4
0.6
0 20 40 60 80 100 120 140
-0.4
-0.2
0
95% percent condence intervals of the estimator for E(S| = w)
(a) = 0.01
100
120
140
160
180
200
0
20
40
60
80
100
0 20 40 60 80 100 120 140
(b) = 0.1
100
120
140
160
180
200
0
20
40
60
80
100
0 20 40 60 80 100 120 140
(c) = 0.5
100
120
140
160
180
200
0
20
40
60
80
100
0 20 40 60 80 100 120 140
(d) = 0.8
100
120
140
160
180
200
0
20
40
60
80
100
0 20 40 60 80 100 120 140
Data Description:
A large North American commercial bank (U.S. Bank)
Analysis is of the Retail Banking service
Period - all weekdays (Monday through Friday) between
January-June, 2006
Observe arrivals between 10:00 and 16:00
Total number of observations is 2,722,129,
out of which 2,683,418 calls where served
Mean Service-Time as a Function of the Waiting-Time
mean service-time - points, tted spline - solid line
190
210
230
250
270
290
150
170
190
210
230
250
270
290
0 50 100 150 200 250 300
Waiting Time
Fitted Spline Curve E(S|>W=w)
Testing the relationship between patience and service-time
Consider only served calls
Omit all non-waiting observations
Divide the observations into 9 groups, by the ranking their
waiting times
Perform a random pairing permutation test (4000 replications)
The value of the original permutation statistic is 46,140.62
A histogram for the distribution of the test statistic
300
400
500
600
700
F
r
e
q
u
e
n
c
y
0
100
200
300
46079 46079.5 46080 46080.5 46081
F
r
e
q
u
e
n
c
y
Statistic's value
In order to estimate E(S| = w) we use the formula
E(S| = w) = g(w)
g
(w)
h
(w)
We t a cubic smooth spline for g(w) - the mean service-time, as
a function of the waiting-time
Designed to handle smooth functions
Enables to simply extract the derivatives
We choose a spline with only 5 knots - smoothness of the derivative
(a) Splines Derivative
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 50 100 150 200 250 300
Waiting Time
Spline's Derivative
(b) Hazard-Rate of the Patience
0.001
0.002
0.003
0.004
0.005
0.006
H
a
z
a
r
d

R
a
t
e
0
0.001
0.002
0.003
0.004
0.005
0.006
0 50 100 150 200 250 300
H
a
z
a
r
d

R
a
t
e
Waiting Time
Kaplan-Meier Estimate Heft Smooth Estimate
Estimator for the mean service-time as a function of the patience of a customer
0
100
200
300
400
500
-200
-100
0
100
200
300
400
500
0 50 100 150 200
Waiting Time
Fitted Spline Curve E(S|T=w)
Future Research
A renement of the estimation procedure of the mean
service-time as a function of the patience
Further research and modeling of the stang rules and
performance measures in the M
t
/GI
/n
t
+ GI queue
Imply the presented model to other databases
Thank You

The Workload Process: Modeling, Inference and Applications: Michael Reich

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

The Workload Process: Modeling, Inference and Applications: Michael Reich

Transféré par

Droits d'auteur :

Formats disponibles

Introduction Relationship Between Service Time and Patience Workload and Oered-Load Empirical Results Future Research

The Workload Process:

- General service-time distribution; May depend on

(w) are the cdf and the hazard-rate function of

- The original samples statistic

(u) [1 G(t u)]du,

(u) [1 G(t u)]du

per d time units

E(S| = t) be of the form:

Vous aimerez peut-être aussi